Generative AI in Enterprises

With all the hype around ChatGPT and other emerging generative AI applications it’s not a surprise that now leaders are looking at how similar technologies can bring value in enterprise environments and if organizations can start using them today.

As always, value comes with the use case and I am sure with the advancement of AI technologies we will see more and more use cases in the enterprise field. But to have a better idea of the usability of such technologies in the enterprise world today, let's first try to understand what sits behind a tool like ChatGPT.

ChatGPT is built on Large Language Models (LLMs). In an essence, LLMs utilize a neural network architecture and is trained on a massive amount of text data (ranging from books and articles to social media posts and online conversations). This is not really news since LLMs have been around since 2018. In the case of ChatGPT we have a tool that automatically generates text based on written prompts in a fashion that is very advanced, creative, and conversational in nature and utilizes internet scale data. No doubt this brings some exciting advancements and opportunities at scale because it is available and accessible to millions of consumers.

However, usually when we talk about enterprise use cases, we need to think about things such as trust, ethics, bias, regulations and to ensure we have guardrails and means for fact-checking. Organizations are usually quite cautious about any chance of publishing incorrect or inappropriate information to customers or users. And while it is impressive and even quite useful to generate an essay and explain a complex topic in understandable fashion to a kid, the very nature of generative AI tools opens the door for the creation of information which may not be factual or fully complete. Or in other words - "best guess" approach is used.

Going back to the use cases of conversational AI that bring value for enterprises, usually they are around understanding, relaying and applying business processes relevant to the organization. And even if we have a model trained really good on massive amounts of general text-based data, it would not “understand” the specific business processes and flows required by specific enterprise. It would simply require training using data that is not available outside of the organization. But even if it is a use case that does not require strictly internal data, usually generally trained language models need some refining or additional training on specific terminology to increase the accuracy to the required level in the enterprise.

Nevertheless, in some cases it could be enough to use generic information to lead the conversation but only until the moment when concrete, factual data is needed to execute a business process. Then the conversational engine would require the ability to ask clarifying questions, disambiguate and understand further details to drive outcomes. Just imagine business processes in banking where specific details are required and no space for "guesses" is allowed. And this leads also to the ability to contextualize and personalize the conversation. In such cases we need "hyper-personalization" so that the specific business process can be led and executed. A language model and text base conversation alone is not enough to execute transactions for examples.

I already mentioned bias which is tightly connected to ethics and in some cases to regulations. And bias can become a real issue if the model is trained on internet scale data which contains inherent human biases. In enterprise environment we need appropriate guardrails to ensure trustworthy AI, to provide transparency and governance and be prepared to explain the outcome of specific AI model (e.g. in case of audit purposes). Obviously, the source of training data is critical, and the quality comes with cost. Some companies may be tempted to use their customers' data to train their models which brings ethical questions and can even lead to loss of competitive advantage for some of these customers.

The hype around ChatGPT is real and it brings excitement about the potential use of similar tools in organizations. Enterprise conversational AI however requires enterprise readiness –  robust foundational capabilities to integrate with existing systems, contextualization and customization capabilities for your own business, and most importantly - AI Governance and controls.

For those of you interested in enterprise ready AI solutions you can check https://www.ibm.com/watson and https://www.ibm.com/artificial-intelligence

#ChatGPT #AIforBusiness #WatsonAssistant #TrustedAI